Upload 2 files

app.py

@@ -0,0 +1,93 @@
+import os
+
+import gradio as gr
+import numpy as np
+import torch
+from lavis.models import load_model_and_preprocess
+from PIL import Image, ImageDraw, ImageFont
+import openai
+
+device = torch.device("cuda") if torch.cuda.is_available() else "cpu"
+
+model, vis_processors, _ = load_model_and_preprocess(
+    name="blip2_opt", model_type="pretrain_opt2.7b", is_eval=True, device=device
+)
+
+openai.api_key = os.environ["OPENAI_API_KEY"]
+
+def generate_caption(image):
+    image = vis_processors["eval"](image).unsqueeze(0).to(device)
+    caption = model.generate({"image": image})
+
+    caption = "\n".join(caption)
+    #use gpt-4 to generate a meme based on the caption
+    response = openai.ChatCompletion.create(
+                model="gpt-4",
+                messages=[
+                    {"role": "system", "content": "Escribe un meme chistoso para una imagen a partir en la descripción dada por el usuario. No uses emojis. El meme tiene que ser corto y gracioso. El output del asistente solo debe ser el meme. Asegúrate que el meme sea tan bueno que se vuelva viral!"},
+                    {"role": "user", "content": caption}
+                ],
+                temperature=0.6
+            )
+    
+    meme_text = response.choices[0].message.content
+    print(meme_text)
+
+    pil_image = Image.fromarray((image.squeeze(0).permute(1, 2, 0).cpu().numpy() * 255).astype(np.uint8))
+
+    # Create a copy of the input PIL image
+    meme_image = pil_image.copy()
+
+    # Convert the copied image to RGBA mode to support transparency
+    meme_image = meme_image.convert('RGBA')
+
+    # Add the text to the image
+    draw = ImageDraw.Draw(meme_image)    
+    
+    # Calculate font size based on text length
+    text_length = len(meme_text)
+    if text_length <= 15:
+        font_size = 24
+    elif text_length <= 30:
+        font_size = 18
+    else:
+        font_size = 14
+
+    font = ImageFont.load_default()
+    text_width, text_height = draw.textsize(meme_text, font=font)
+    
+    # Calculate the position to place the text at the top and center horizontally
+    x = (meme_image.width - text_width) // 2
+    y = 10  # Adjust this value to add more or less padding from the top
+
+    draw.text((x, y), meme_text, fill=(255, 255, 255), font=font)
+
+    # Convert the meme_image back to RGB mode
+    meme_image = meme_image.convert('RGB')
+
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    return meme_image
+
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+        "### Memero - Generador de Memes"
+    )
+    gr.Markdown(
+        "Escribe un meme en español a partir de una imagen."
+    )
+
+    with gr.Row():
+        with gr.Column():
+            input_image = gr.Image(label="Imagen", type="pil")
+            btn_caption = gr.Button("Generar meme")
+            output_text = gr.Image(label="Meme", lines=5)
+
+    btn_caption.click(
+        generate_caption, inputs=[input_image], outputs=[output_text]
+    )
+
+demo.launch()

requirements.txt

@@ -0,0 +1,4 @@
+torch
+torchvision
+salesforce-lavis
+openai